search

TheSouthFrog / stylealign

[ICCV 2019]Aggregation via Separation: Boosting Facial Landmark Detector with Semi-Supervised Style Transition
182 stars 28 forks source link

Hi！some question about re-training your pytorch code. #15

Open wqz960 opened 4 years ago

wqz960 commented 4 years ago

Hi! Thanks for your wonderful code. First, In your pytorch code, the size of input images is 1281283, have you tested for 2562563, can the network provide similar result on different input size? Second, all face image should be aligned?
Third, I downloaded your wflw-full dataset, how to choose the K sample image for style transform? Thank you!

TheSouthFrog commented 4 years ago

  1. Yes, I think the results should be very similar. You may try to deepen the network structures and are expected to see similar results.

  2. Right. Actually the images I provided are aligned.

  3. Images in wflw-full are all generated samples, with 8 times larger then the original dataset, which means I choose k equals to 8 here. That's just an example for you to train a facial landmark detector without training the generator. You can choose to add the generated data and original data together to train your detector.

wqz960 commented 4 years ago

Thank you for your kind help! Can you provide me some network structures for 256 size, the vunet is open-source for tensorflow, it is too difficult for me to re-implement on pytorch. Thank you!